Lazyweb Ab Test Research

Lazyweb A/B Test Research

Use Lazyweb mobile-only experiment evidence to answer growth PM questions. The public gateway and the richer backend/internal MCP surfaces are not identical, so start from the live tool schema before choosing how to retrieve evidence.

MCP Setup

Use hosted Lazyweb MCP tools for all database-backed evidence. First list the available tools and run lazyweb_health.

• lazyweb_health — verify Lazyweb MCP connectivity.
• lazyweb_search_ab_tests — current public mobile-only gateway for A/B Test Agent research, included free.
• lazyweb_search — pull visual design references to pair with experiment evidence.
• lazyweb_compare_image / lazyweb_find_similar — visual reference retrieval when the target screen or adjacent examples would clarify the recommendation.
• lazyweb_list_categories — public category browsing helper.

Pass skill: "ab-test-research" on every call. Include "skill": "ab-test-research" in the arguments of each lazyweb_* tool call — for example {"query": "pricing page", "limit": 30, "skill": "ab-test-research"}. This is optional analytics metadata Lazyweb uses to understand which skills are used; never drop or change a real argument for it.

Also pass version: "<x.y.z>" on every call. Read ~/.lazyweb/VERSION once per session at skill start (e.g. cat "$HOME/.lazyweb/VERSION" 2>/dev/null || echo 0.0.0); fall back to "0.0.0" if the file is missing or unreadable — never block on this. Include "version": "<that-value>" in the arguments of every lazyweb_* tool call alongside the existing skill arg — for example {"query": "pricing page", "limit": 30, "skill": "ab-test-research", "version": "0.4.5"}. Optional analytics metadata Lazyweb uses to track which skill-pack versions are running; never drop or change a real argument for it.

If Lazyweb MCP is not installed or auth fails, tell the user: "Lazyweb MCP is not installed. Run curl -fsSL https://www.lazyweb.com/install.sh | bash, reload this client, then rerun this skill." Then continue with general web research only if the user wants a degraded fallback.

Current public lazyweb_search_ab_tests arguments:

{
  "target_screen_description": "trial reminder onboarding paywall",
  "product": "Example App",
  "category": "Health & Fitness",
  "conversion_goal": "trial start rate",
  "constraints": "keep annual plan visible",
  "operation": "research",
  "experiment_ids": ["exp_123"],
  "include_images": true,
  "target_image_url": "https://example.com/screen.png",
  "limit": 25,
  "analysis_experiment_limit": 10,
  "visual_inspection_budget": 0,
  "interesting_learning": false,
  "high_design_bar": false
}

The public A/B wrapper is included free. If lazyweb_search_ab_tests is available, call it directly and use the returned experiment evidence. If the tool is unavailable or returns no matching experiments, say that experiment evidence was unavailable for this query, then continue with Lazyweb visual references when useful.

category is the public gateway's industry filter. product is forwarded as target context for the user's app, not as a reason to force exact company matching; do not retry exact product/company spellings or trust a zero-result response when warnings indicate a product/company filter was applied. If the product is useful context, include it, but make the retrieval query screen-pattern plus industry led.

Backend/Internal Experiment Tools

Some backend or internal MCP surfaces expose these richer generic experiment tools. Use them only when the current tool list includes them:

• lazyweb_find_experiments — retrieve generic _experiments evidence.
• lazyweb_recent_experiments — retrieve the latest 10, 25, or 50 _experiments rows.
• list_companies_by_categories — turn category names into company IDs.

_experiments is a limited screenshot-diff evidence set. It is generic across screens and categories, not paywall-only. Treat learning text as directional hypotheses, not statistically measured lift.

Full lazyweb_find_experiments filter matrix:

{
  "query": "trial reminder onboarding upsell",
  "company": "Example App",
  "category": "Health & Fitness",
  "screen_type": "onboarding upsell",
  "platform": "mobile",
  "company_ids": [123, 456],
  "canonical_ids": [789],
  "since_iso": "2026-06-01T00:00:00Z",
  "limit": 50,
  "app_store_rank_max": 50,
  "app_store_overall_rank_max": 50,
  "app_store_category_rank_max": 25,
  "high_design_bar": true
}

Full lazyweb_recent_experiments filter matrix:

{
  "limit": 25,
  "company": "Example App",
  "category": "Health & Fitness",
  "platform": "mobile",
  "company_ids": [123, 456],
  "app_store_rank_max": 50,
  "app_store_overall_rank_max": 50,
  "app_store_category_rank_max": 25,
  "high_design_bar": true
}

lazyweb_search_ab_tests exposes interesting_learning and high_design_bar. Leave interesting_learning as false by default. Set it to true only when the user explicitly asks for uncommon, surprising, or contrarian learnings; clearly label those as limited evidence. Set high_design_bar only when the user asks for premium, stronger, high-design-bar, or best-designed examples.

Do not route through legacy paywall-specific research tools. If a paywall appears in the evidence, treat it as one screen type among many.

Workflow

1. Ground the product question. Identify product/app, category, screen or flow, platform, target metric, and constraints.

2. Choose the available evidence path.

   • If the current MCP surface only exposes the public gateway, call lazyweb_search_ab_tests.
   • If lazyweb_find_experiments is exposed, retrieve generic experiment rows with the strongest filters available.
   • If the user asks for recent/latest tests and lazyweb_recent_experiments is exposed, use that tool with a limit of 10, 25, or 50.
   • If list_companies_by_categories is exposed and the category is known, call it first and pass the returned company_ids into lazyweb_find_experiments.

Public gateway example:

{
  "target_screen_description": "trial reminder onboarding upsell",
  "product": "Example App",
  "category": "Health & Fitness",
  "conversion_goal": "trial start rate",
  "limit": 25,
  "analysis_experiment_limit": 10
}

Backend/internal retrieval example:

{
  "query": "trial reminder onboarding upsell",
  "category": "Health & Fitness",
  "screen_type": "onboarding upsell",
  "company_ids": [123, 456],
  "limit": 30
}

Use minimal filters for popular apps or broad best-practice questions. On the public gateway, prefer screen-pattern plus category for industry context; keep product as target context only, not as an exact company filter. Use richer filters only on backend/internal surfaces whose live schema exposes them.

When the user asks for high-design-bar companies, premium examples, best-designed apps, or stronger taste filtering, add this only to tools whose live schema exposes it:

{"high_design_bar": true}

This filters to companies where companies.high_design_bar = true on the backend/internal surfaces that support it.

For "recent", "latest", or "what changed lately" requests, call lazyweb_recent_experiments when it is exposed, with limit set to 10, 25, or 50:

{"limit": 25}

For ranked App Store slices, add rank filters:

{
  "category": "Health & Fitness",
  "app_store_overall_rank_max": 50,
  "app_store_category_rank_max": 25,
  "limit": 25
}

3. Supplement with design references. Call lazyweb_search for the same screen or flow when visual examples would make the recommendation clearer. Read visionDescription before relying on any screenshot, and embed returned optimized imageUrl values directly instead of downloading Lazyweb images locally. Never repeat an identical query — page deeper with offset and follow pagination.next_offset; on no_matches/low_coverage warnings use the closest result or note the gap instead of rephrasing in a loop, and on company_not_in_library use a suggested company or drop the filter.

4. Synthesize like a growth PM. Answer with:

   • Relevant observed experiments and what changed.
   • Likely hypothesis behind each change.
   • Target metric and guardrail metric.
   • Recommended test sequence.
   • Evidence strength and gaps.
   • Where the user should not overgeneralize.

5. Be honest about weak evidence. If the A/B wrapper is unavailable, or the backend/internal retrieval tools return few or weak matches, say that directly and fall back to general best practices only after labeling them as inference.

Output Shape

For a quick strategy question, answer in chat. For anything the user will act on, write a durable HTML report to .lazyweb/ab-test-research/{topic}-{date}/report.html. The report must center the actual experiments — control vs variant, what changed, and the learning — not just a synthesized opinion.

Publish a Shareable Link (whenever an HTML report was written)

Every report is auto-published to lazyweb.com so the user can share it with teammates. Right after writing report.html, run this with $REPORT_DIR set to .lazyweb/ab-test-research/{topic}-{date}:

IDEMPOTENCY_KEY="${REPORT_DIR#.lazyweb/}"   # stable per-report key (e.g. ab-test-research/{topic}-{date}); send the SAME value every attempt so retries dedupe to one link
LAZYWEB_TOKEN=$(cat "$HOME/.lazyweb/lazyweb_mcp_token" 2>/dev/null || true)
if [ -n "$LAZYWEB_TOKEN" ]; then
  # Tier 1 - local install: direct POST (idempotency_key dedupes a re-run)
  python3 - "$REPORT_DIR" "$LAZYWEB_TOKEN" "ab-test-research" "$IDEMPOTENCY_KEY" <<'PUBLISH_EOF'
import base64, json, pathlib, sys, urllib.error, urllib.request
report_dir, token, skill, idem = pathlib.Path(sys.argv[1]), sys.argv[2], sys.argv[3], sys.argv[4]
version_file = pathlib.Path.home() / ".lazyweb" / "VERSION"
version = version_file.read_text().strip() if version_file.exists() else "0.0.0"
html = (report_dir / "report.html").read_text(encoding="utf-8")
refs = report_dir / "references"
assets = [
    {"name": p.name, "b64": base64.b64encode(p.read_bytes()).decode()}
    for p in (sorted(refs.iterdir()) if refs.is_dir() else [])
    if p.is_file()
]
body = json.dumps({"skill": skill, "version": version, "html": html, "assets": assets, "idempotency_key": idem}).encode()
req = urllib.request.Request(
    "https://www.lazyweb.com/api/reports",
    data=body,
    headers={"Content-Type": "application/json", "Authorization": f"Bearer {token}"},
)
try:
    resp = json.loads(urllib.request.urlopen(req, timeout=90).read())
    print(f"SHAREABLE_URL: {resp['url']}")
except urllib.error.HTTPError as exc:
    print(f"PUBLISH_FAILED: {exc.code} {exc.read().decode()[:500]}")
except Exception as exc:
    print(f"PUBLISH_SKIPPED: {exc}")
PUBLISH_EOF
else
  # Tier 2 - no local token (hosted/cloud agent): publish via the MCP tool (see below)
  echo "PUBLISH_VIA_MCP_TOOL idempotency_key=$IDEMPOTENCY_KEY report_dir=$REPORT_DIR"
fi

Exactly one tier runs - never both.

• Tier 1 SHAREABLE_URL: - include the link: "Shareable link: {url} (unlisted - anyone with the link can view)".
• Tier 1 PUBLISH_FAILED: 400 ... - the body names what is unhostable (e.g. missing_assets). Fix the report and re-run the publish ONCE.
• Tier 1 PUBLISH_SKIPPED: - say nothing; the local report stands (the user has the file).
• Tier 2 PUBLISH_VIA_MCP_TOOL ... - you have no local token (hosted session), so publish with the Lazyweb MCP tool instead:
  1. Size-check first: if report.html plus the references/ files together exceed ~7MB, do NOT call the tool - tell the user the report was too large to publish from a hosted session (it is saved locally) and stop.
  2. Otherwise call lazyweb_publish_report with: html = the contents of report.html; assets = each references/ file as {"name": <filename>, "b64": <base64 of the bytes>}; report_skill = "ab-test-research"; idempotency_key = the value printed after idempotency_key=.
  3. On { ok: true, url } -> show "Shareable link: {url} (unlisted - anyone with the link can view)".
  4. On { ok: false } -> tell the user publishing failed and why (the error field); the report is saved locally. If code is REPORT_VALIDATION_ERROR and detail names missing assets, fix and call ONCE more; otherwise do not retry. Unlike Tier 1, do NOT stay silent on a Tier-2 failure - a hosted user has no local file to fall back on, so they need the link or the reason.

Hosting-safe HTML (the template already complies — keep it that way)

The hosted copy is served byte-for-byte, so the report must only use:

• inline CSS and inline <script> — never an external <script src=...>
• images via the absolute optimized imageUrl/image_url URLs Lazyweb returns, or relative references/{filename} paths for locally saved screenshots
• no file:// URLs and no absolute local paths (/Users/..., C:\...)

Fields lazyweb_search_ab_tests (operation research) returns by default

Per experiment in evidence.experiments[] (no flags needed):

• company.company_name, company.category, company.subcategory, company.app_store_ranking
• control.{imageUrl, image_url, path, vision_description} and variant.{imageUrl, image_url, path, vision_description}
• what_changed (text: the concrete control→variant diff), learning (text: directional hypothesis + why), evidence_confidence, platform, experiment_id, target_screen_description

Top level: recommendations[] (each cites an experiment_id + target_metric + guardrail_metric + confidence), strong_points, weak_points, dataset_caveat.

Experiment images are returned as optimized URLs. Use control.imageUrl or control.image_url, and variant.imageUrl or variant.image_url, directly. Some adjacent experiment objects may expose aliases such as control_image_url, controlImageUrl, variant_image_url, or variantImageUrl; use those directly when present. Supabase storage-backed URLs are signed for 365 days. Do not use screenshot IDs, and do not construct storage URLs from raw path values. If an image URL is missing, drop that <img> and keep the vision_description. company_name is a crawl seed — you may clean an obvious slug but never invent a brand; flag any /figma/ or !-prefixed path as a non-production capture in the caption.

(Visual refs from lazyweb_search also embed via their returned imageUrl/image_url fields.)

Content outline (semantic HTML, LIGHT theme)

# A/B Test Research: {Flow / Question}

## Agent Instructions            (section #1 — see Report essentials below)

## Recommendations               (ranked TABLE: Recommendation · Target metric · Guardrail · Confidence · Rests on exp_id(s).
                                  Rows map to recommendations[]; every row cites the experiment_id(s) it rests on.)

## The A/B tests   ← REQUIRED, the centerpiece
{One card per experiment from evidence.experiments[] (see card markup below).
 Show control vs variant images + what_changed + learning + confidence.
 Render dataset_caveat ONCE, directly under this heading. The exp ids here match the
 "Rests on" column above so a reader can jump from a recommendation to its evidence.}

## Strong / Weak points          (short bullets straight from strong_points[] / weak_points[])

## Where not to overgeneralize   (the corpus is mobile-subscription-centric; learnings are
                                  directional screenshot-diff signals, not measured lift)

## References (optional)         (only if lazyweb_search refs were pulled; `.deck` snap-carousel, scroll-snaps with ◀ ▶ prev/next buttons)

"The A/B tests" card (per experiment)

<section id="ab-tests">
  <h2>The A/B tests</h2>
  <p class="caveat">{dataset_caveat verbatim}</p>

  <article class="exp">
    <header class="exp-h"><span class="exp-co">{company.company_name}</span>
      <span class="exp-meta">{category} › {subcategory} · Rank #{app_store_ranking} · {platform}</span>
      <span class="exp-id">exp {experiment_id}</span></header>
    <p class="exp-target">Target screen: {target_screen_description}</p>
    <div class="flip">
      <figure><img src="{control.imageUrl or control.image_url}" alt="Control — {company}" loading="lazy"
       
        <figcaption>Control<span class="vd">{control.vision_description}</span></figcaption></figure>
      <figure><img src="{variant.imageUrl or variant.image_url}" alt="Variant — {company}" loading="lazy"
       
        <figcaption>Variant<span class="vd">{variant.vision_description}</span></figcaption></figure>
    </div>
    <dl class="exp-facts"><dt>What changed</dt><dd>{what_changed}</dd>
      <dt>Learning</dt><dd>{learning}</dd><dt>Confidence</dt><dd>{evidence_confidence}</dd></dl>
  </article>
  <!-- repeat <article class="exp"> per experiment -->
</section>

Report essentials (apply to the report you write)

A. Agent Instructions — report section #1

The report opens with an Agent Instructions callout: one plain human sentence, then a copy-pastable block written FOR A DOWNSTREAM CODING AGENT. Emit exactly this structure:

<section id="agent-instructions" class="agent-instructions">
  <div class="ai-head"><span class="ai-badge">FOR THE CODING AGENT</span>
    <button class="ai-copy" type="button"
      var sec=this.closest('.agent-instructions'); var txt=sec.querySelector('.ai-block').innerText;
      var done=function(ok){this.textContent=ok?'Copied':'Press Cmd/Ctrl+C';setTimeout(function(){this.textContent='Copy';}.bind(this),1500);}.bind(this);
      if(navigator.clipboard&&navigator.clipboard.writeText){navigator.clipboard.writeText(txt).then(function(){done(true);},function(){done(false);});}
      else{var r=document.createRange();r.selectNodeContents(sec.querySelector('.ai-block'));var s=getSelection();s.removeAllRanges();s.addRange(r);try{document.execCommand('copy');done(true);}catch(e){done(false);}}">Copy</button>
  </div>
  <p class="ai-human">{one human sentence: the single most important test to run first}</p>
  <pre class="ai-block">{COPY BLOCK — fill the braces from this report}</pre>
</section>

Copy-block text (keep these exact labels; fill {REPORT_PATH} with the absolute path of the report.html you wrote):

LAZYWEB REPORT — AGENT HANDOFF
Use the report at {REPORT_PATH} as a starting point for {TASK}.

TOP RECOMMENDATIONS (do first):
1. {rec 1, one imperative line}
2. {rec 2}
3. {rec 3}

INDEX ON: {1-3 best-evidenced experiment learnings}
DO NOT OVER-INDEX ON: {directional-not-measured learnings, off-category experiments, single-experiment signals}
DIVE FURTHER: {next Lazyweb skill or MCP tool} — {why}

Evidence basis: A/B experiments (screenshot-diff) · {DATE}

For THIS skill, {TASK} = "prioritizing and shipping {flow} experiments grounded in what comparable apps have already tested", and DIVE FURTHER → "/lazyweb-optimize-paywall to turn a paywall learning into a falsifiable redesign, or lazyweb_search_ab_tests operation=grab with the cited experiment_id(s)".

B. Conciseness & "show, don't tell"

No length target — let the evidence set the length. Lead with value (Agent Instructions + the ranked Recommendations table). Show, don't tell: the centerpiece is the real control/variant screenshots, not prose. Index every recommendation on a named experiment_id + its learning, never generic growth-speak. If you must illustrate a proposed variant layout that no screenshot shows, use the mock-frame component (mobile/desktop) — never ASCII art.

C. HTML requirements (LIGHT theme — match lazyweb-deep-design-research)

• Single HTML file, inline CSS (no external CSS/JS dependencies; the one inline onclick copy handler above is allowed).
• Light/white design — do NOT use a dark/black theme. System fonts, max-width:900px, white background, comfortable line-height, light borders, #f6f8fa table headers.
• Use the shared design tokens and include the CSS below in <style>.
• Agent Instructions is the first section, styled as the light-blue callout. The "A/B tests" section is required whenever ≥1 experiment is returned and sits right after Recommendations.
• dataset_caveat appears once. Use returned experiment image URL fields directly; degrade gracefully via the onerror hook.
• Open in the browser: open "$REPORT_DIR/report.html".

:root{--ink:#1f2328;--mut:#57606a;--line:#d0d7de;--soft:#eef4fb;--accent:#0969da}
body{font:16px/1.6 -apple-system,BlinkMacSystemFont,"Segoe UI",Helvetica,Arial,sans-serif;color:var(--ink);background:#fff;max-width:900px;margin:0 auto;padding:40px 22px}
table{border-collapse:collapse;width:100%;font-size:14px}th,td{border:1px solid var(--line);padding:7px 9px}th{background:#f6f8fa;text-align:left}
.agent-instructions{background:var(--soft);border-left:4px solid var(--accent);border-radius:8px;padding:14px 16px;margin:18px 0}
.ai-head{display:flex;justify-content:space-between;align-items:center;gap:12px;margin-bottom:8px}
.ai-badge{font-size:11px;font-weight:700;letter-spacing:.04em;color:#0a3b78}
.ai-copy{font:600 12px/1 inherit;cursor:pointer;border:1px solid var(--accent);color:var(--accent);background:#fff;border-radius:6px;padding:5px 11px}.ai-copy:hover{background:var(--accent);color:#fff}
.ai-human{margin:0 0 10px;font-size:15px}
.ai-block{white-space:pre-wrap;word-break:break-word;background:#fff;border:1px solid var(--line);border-radius:6px;padding:12px 13px;margin:0;font:13px/1.5 ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;color:var(--ink);user-select:all}
#ab-tests .caveat{font-size:13px;color:var(--mut);background:#fff8e6;border:1px solid #f0e0b0;border-radius:6px;padding:8px 11px;margin:8px 0 16px}
.exp{border:1px solid var(--line);border-radius:10px;padding:14px 16px;margin:14px 0;background:#fff}
.exp-h{display:flex;flex-wrap:wrap;gap:8px;align-items:baseline}.exp-co{font-weight:700}.exp-meta{color:var(--mut);font-size:13px;flex:1}
.exp-id{font:12px ui-monospace,Menlo,monospace;color:var(--mut);background:#f6f8fa;border-radius:5px;padding:1px 7px}
.exp-target{font-size:13px;color:var(--mut);margin:6px 0 10px}
.exp-pair{display:grid;grid-template-columns:1fr 1fr;gap:12px}@media(max-width:600px){.exp-pair{grid-template-columns:1fr}}
.exp-pair img{width:100%;height:auto;max-height:580px;object-fit:contain;border:1px solid var(--line);border-radius:8px;background:#fafbfc}
.exp-pair figcaption{font-size:12px;font-weight:600;margin-top:5px}.exp-pair .vd{display:block;font-weight:400;color:var(--mut);font-size:11.5px;margin-top:2px}
.exp-pair figure.img-missing img{display:none}
.exp-pair figure.img-missing figcaption::after{content:" — image unavailable; see description";color:#cf222e;font-weight:400}
.exp-facts{display:grid;grid-template-columns:max-content 1fr;gap:4px 12px;margin:12px 0 0;font-size:14px}.exp-facts dt{font-weight:600;color:var(--mut)}
.mock{margin:14px 0}.mock .frame{border:1px solid var(--line);border-radius:14px;background:#fff;overflow:hidden}.mock.mobile .frame{max-width:300px;border-radius:26px;border:8px solid #1f2328}.mock.desktop .frame{max-width:760px}
.mock .body{padding:14px;display:flex;flex-direction:column;gap:10px}.mock .box{background:var(--soft);border:1px dashed #b9c7d6;border-radius:8px;min-height:34px;display:flex;align-items:center;justify-content:center;color:#4a5a6a;font-size:12px;padding:8px}.mock .box.cta{background:var(--accent);border:0;color:#fff;font-weight:600}.mock .row{display:flex;gap:10px}.mock .row>.box{flex:1}.mock .cap{font-size:12px;color:var(--mut);margin-top:6px;text-align:center}

Operating principles & evidence components (REQUIRED - overrides convenience)

Operating principles (apply to every report you write)

These four rules override convenience. A report that breaks them is non-conforming, even if every section is present.

1. Show, don't tell — every claim carries its proof. Any assertion — a pattern, anti-pattern, idea, hypothesis, "what's working" item, convention check, recommendation, or A/B learning — must render the real screenshot(s) or experiment that demonstrate it, beside the claim, never a scroll away. When more than one reference backs a claim, render them as a .deck snap-carousel — a horizontally-scrolling strip that snaps card-to-card with ◀ ▶ prev/next buttons — so the reader can step through the proof, never hidden behind a prose list or a "see Section 5" pointer. Prevalence words ("most", "near-universal", "dominant") must be backed by a shown count ("5 of 9 references"), never an adjective alone.

2. Be opinionated; carry the decision. Lead with ONE ranked recommended path, marked as the lead pick (.lead ribbon) in the human-visible body — not only in the agent copy block. Tag every other option Do / Explore / Skip (or P0/P1/P2) with a one-line "skip if". No ties among top picks; no flat undifferentiated menu. The "Skip" rows must link to the evidence (e.g. the anti-pattern screenshot) so the skip decision is shown, not just asserted.

3. Maximize confidence with evidence + data. Back each recommendation with what worked for OTHER apps (real screenshots) PLUS supporting data: a prevalence count across the corpus ("seen in N of M examples") and, where the screen is growth/monetization, A/B experiment evidence via lazyweb_search_ab_tests. If no experiment data exists, say so explicitly ("no experiment data found — recommendation is design-prevalence-based") and substitute the prevalence count as the directional signal. Never let a recommendation render with neither a visual nor a number behind it.

4. Be truth-seeking — never overclaim. Label evidence strength honestly with an .ebadge on every claim/card/rec: Measured (real lift number) vs Directional (screenshot-diff / visual prevalence, no lift) vs Single-source / Off-category. Forbid comparative-performance verbs ("outperforms", "underperforms") unless a measurement backs them. Put a one-line corpus-strength banner (.corpus) right after Agent Instructions when evidence is single-source, thin, or context-mismatched. Tag any reference whose brand was inferred from a URL/vision-description ("brand inferred — verify"). Show absence claims with evidence-of-search (queries run × screens reviewed + the closest near-miss). Never invent a reference, a metric, or a company name. Never use ASCII/box-drawing <pre> art for a layout — render the .mock mock-frame or a generated image.

Evidence components (render claims with these — never prose alone)

Put the CSS below in the report <style> and adapt the markup per claim. Reuse the existing tokens (--ink:#1f2328; --mut:#57606a; --line:#d0d7de; --soft:#eef4fb; --accent:#0969da).

• Patterns, Anti-Patterns (use .tag.avoid), Unique Angles, "What's working", Convention-check Missing/Unusual cells, hypothesis "supporting evidence", brainstorm "Applied here" -> render as .pat cards: name + verdict .tag + .prev count ("seen in N of M references") + one-line claim + a .deck snap-carousel of 2-4 real screenshots that exhibit it (caption = company + the exact UI detail from visionDescription). The proof sits WITH the claim, never a scroll away. Show the pattern; don't just list brand names.
• The decision section (deep-design-research "What to build first", lite-design-research "Recommended path", brainstorm "Which ideas to prototype", design-improve idea order, ab-test "Recommendations", optimize-paywall "Prioritization") -> render a Decision legend (.legend, one .legend-row per rec in rank order — the whole ranking graspable in one glance) ABOVE a .recs stack of recommendation cards (.rec; the #1 card is .rec.lead with the START-HERE hero + browser-chrome frame). Each card carries a BIG inline-legible proof (desktop 16:10 above-the-fold crop / mobile whole-screen via .recs.mobileset) — never a tiny click-into thumbnail — plus a .verdict (Do/Explore/Skip), an .ebadge evidence label, a prevalence count, and a labeled .skiprow. NO table; exactly one lead; no ties.
• Honesty -> put a .corpus banner right after Agent Instructions (basis / breadth / count / confidence); put an .ebadge (Measured | Directional | Single-source-or-off-category) on every claim, card, and recommendation. Never echo a raw "high"; never use "outperforms/underperforms" without a measured lift; tag crawl-seed/URL-inferred brands "brand inferred - verify".
• Control vs variant (A/B) -> use the .flip two-up grid (stacks on narrow screens; scroll-snaps with ◀ ▶ prev/next buttons); if vision_description is empty, synthesize the caption from what_changed and tag it "agent-described".

Every embedded screenshot must be FULLY VISIBLE and legible inline — never require a click to view, never a tiny thumbnail-as-proof, never a weird letterboxed ratio. Mobile/portrait shots: show the WHOLE screen (default .deck figure, no crop). Desktop/long-scroll pages: tag the figure .shot-web to crop to above-the-fold (16:10 from the top) at a size large enough to understand on its own. No "open full image" links. A report that asserts a claim with neither a visual nor a number behind it — or that hides its proof behind a click or a sliver crop — is non-conforming.

.deckwrap{margin:8px 0}
.deck{display:flex;gap:12px;overflow-x:auto;scroll-snap-type:x mandatory;-webkit-overflow-scrolling:touch;padding:4px 2px 10px;scrollbar-width:thin}
.deck>figure{flex:0 0 86%;max-width:320px;scroll-snap-align:center;margin:0;border:1px solid var(--line);border-radius:10px;overflow:hidden;background:#fff}
@media(min-width:620px){.deck>figure{flex-basis:46%}}
.deck.web>figure,.deck>figure.shot-web{flex-basis:92%;max-width:560px}
@media(min-width:620px){.deck.web>figure,.deck>figure.shot-web{flex-basis:60%}}
.deck-nav{display:flex;gap:6px;justify-content:flex-end;margin-top:2px}
.deck-nav button{cursor:pointer;border:1px solid var(--line);background:#fff;color:var(--ink);border-radius:6px;width:34px;height:28px;font-size:13px;line-height:1}
.deck-nav button:hover{background:var(--soft);border-color:var(--accent);color:var(--accent)}
.deck>figure>img{display:block;width:100%;height:auto;max-height:620px;object-fit:contain;background:#fafbfc}
.deck>figure.shot-web>img{aspect-ratio:16/10;max-height:none;object-fit:cover;object-position:top}
.deck>figure.tall>img{object-fit:contain}
.deck>figure.img-missing>img{display:none}
.deck>figure.img-missing figcaption::after{content:' — image unavailable; see description';color:#cf222e}
.deck .cap{font-size:12px;color:var(--mut);padding:7px 9px;line-height:1.4}
.deck .cap b{color:var(--ink)}
.deck .src{font-size:10.5px;font-weight:700;letter-spacing:.03em;color:var(--accent)}
.deck-hint{font-size:11.5px;color:var(--mut);margin:-2px 0 6px}
.pat{border:1px solid var(--line);border-radius:12px;padding:14px 16px;margin:14px 0;background:#fff}
.pat-h{display:flex;flex-wrap:wrap;align-items:baseline;gap:8px;margin-bottom:4px}
.pat-h h3{margin:0;font-size:16px}
.pat-claim{color:var(--mut);font-size:14px;margin:2px 0 10px}
.prev{font:600 11.5px/1 inherit;color:#0a3b78;background:var(--soft);border-radius:20px;padding:4px 10px}
.tag{font:700 10.5px/1 inherit;letter-spacing:.03em;border-radius:5px;padding:3px 7px}
.tag.strong{color:#0a5d2a;background:#e6f4ea;border:1px solid #b7e0c4}
.tag.directional{color:#8a5a00;background:#fff8e6;border:1px solid #f0e0b0}
.tag.weak{color:#6e7781;background:#f6f8fa;border:1px solid var(--line)}
.tag.avoid{color:#a40e26;background:#fdeef0;border:1px solid #f5c2c7}
:root{
  --ink:#1f2328; --mut:#57606a; --line:#d0d7de; --soft:#eef4fb; --accent:#0969da;
  --do-fg:#0a5d2a; --do-bg:#e6f4ea; --do-bd:#b7e0c4;
  --ex-fg:#8a5a00; --ex-bg:#fff8e6; --ex-bd:#f0e0b0;
  --sk-fg:#6e7781; --sk-bg:#f6f8fa; --sk-bd:#e3e7eb;
  --single-fg:#a40e26; --single-bg:#fdeef0; --single-bd:#f5c2c7;
}

/* ===== GRAFT 1 — DECISION LEGEND (whole ranking at a glance; degrades to anchors) ===== */
.legend{border:1px solid var(--line);background:#fff;border-radius:12px;padding:6px;margin:0 0 26px;overflow:hidden}
.legend-row{display:grid;grid-template-columns:34px minmax(0,1fr) auto auto;align-items:center;gap:12px;padding:9px 10px;border-radius:8px;text-decoration:none;color:inherit}
.legend-row + .legend-row{border-top:1px solid #eef1f4}
.legend-row:hover{background:var(--soft)}
.legend-row.is-lead{background:linear-gradient(0deg,#fff,var(--soft))}
.lg-rank{font:800 14px/1 inherit;text-align:center;width:30px;height:30px;border-radius:50%;display:flex;align-items:center;justify-content:center;color:#fff;background:var(--mut)}
.legend-row.is-lead .lg-rank{background:var(--accent)}
.lg-rank.r2{background:#3f6896}.lg-rank.r3{background:#6b7787}.lg-rank.r4{background:#8c96a1}
.lg-name{font-weight:650;min-width:0}
.lg-name .lg-why{display:block;font-weight:400;color:var(--mut);font-size:12px;line-height:1.35;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
.lg-ev{font:700 10px/1 inherit;color:var(--mut);white-space:nowrap;letter-spacing:.02em}

/* ===== shared verdict chip + evidence badge (reused everywhere) ===== */
.verdict{font:700 10.5px/1 inherit;border-radius:6px;padding:5px 9px;white-space:nowrap;letter-spacing:.03em;text-transform:uppercase;border:1px solid transparent;display:inline-flex;align-items:center;gap:5px}
.verdict .ico{font-size:10px;line-height:1}
.verdict.do{color:var(--do-fg);background:var(--do-bg);border-color:var(--do-bd)}
.verdict.explore{color:var(--ex-fg);background:var(--ex-bg);border-color:var(--ex-bd)}
.verdict.skip{color:var(--sk-fg);background:var(--sk-bg);border-color:var(--sk-bd)}
.ebadge{font:700 10px/1 inherit;letter-spacing:.02em;border-radius:20px;padding:5px 9px;display:inline-flex;align-items:center;gap:6px;white-space:nowrap}
.ebadge .dot{width:7px;height:7px;border-radius:50%;display:inline-block}
.ebadge.strong{color:var(--do-fg);background:var(--do-bg);border:1px solid var(--do-bd)}.ebadge.strong .dot{background:var(--do-fg)}
.ebadge.directional{color:var(--ex-fg);background:var(--ex-bg);border:1px solid var(--ex-bd)}.ebadge.directional .dot{background:var(--ex-fg)}
.ebadge.single{color:var(--single-fg);background:var(--single-bg);border:1px solid var(--single-bd)}.ebadge.single .dot{background:var(--single-fg)}
.caveat-inline{font:700 9.5px/1 inherit;color:var(--ex-fg);background:var(--ex-bg);border:1px solid var(--ex-bd);border-radius:5px;padding:3px 6px;margin-left:2px;white-space:nowrap}

/* ===== RECOMMENDATION CARDS (replaces .ranked table) ===== */
.recs{display:flex;flex-direction:column;gap:16px}
.rec{position:relative;display:grid;grid-template-columns:minmax(0,1.18fr) minmax(0,1fr);background:#fff;border:1px solid var(--line);border-radius:14px;overflow:hidden;box-shadow:0 1px 3px rgba(31,35,40,.05)}

/* GRAFT 2 — LEAD HERO treatment for #1 (replaces .lead callout) */
.rec.lead{border:2px solid var(--accent);box-shadow:0 6px 22px rgba(9,105,218,.14);background:linear-gradient(0deg,#fff,var(--soft));margin-top:11px;grid-template-columns:minmax(0,1.32fr) minmax(0,1fr)}
.rec.lead::before{content:'\2605 RECOMMENDED PATH \2014 START HERE';position:absolute;z-index:4;top:-11px;left:18px;background:var(--accent);color:#fff;font:700 10px/1 inherit;letter-spacing:.05em;border-radius:10px;padding:6px 11px;box-shadow:0 2px 6px rgba(9,105,218,.35)}

/* proof side */
.rec-proof{position:relative;background:#0d1117;min-width:0;line-height:0}
.rec-proof .frame{display:block;height:100%;width:100%;border-right:1px solid var(--line);position:relative}
.browserbar{display:flex;align-items:center;gap:6px;height:30px;padding:0 11px;background:#f6f8fa;border-bottom:1px solid var(--line);position:absolute;top:0;left:0;right:0;z-index:2;line-height:1}
.browserbar i{width:9px;height:9px;border-radius:50%;background:#d0d7de;display:block}
.browserbar i:nth-child(1){background:#ff5f57}.browserbar i:nth-child(2){background:#febc2e}.browserbar i:nth-child(3){background:#28c840}
.browserbar .url{margin-left:8px;font:600 10.5px/1 ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--mut);background:#fff;border:1px solid var(--line);border-radius:20px;padding:5px 12px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;flex:1;min-width:0}
/* desktop proof: 16:10 above-the-fold crop — plain <img>, legible with JS off */
.rec-proof img{display:block;width:100%;height:100%;aspect-ratio:16/10;object-fit:cover;object-position:top;background:#161b22}
.rec.lead .rec-proof img{aspect-ratio:16/9}
.rank-badge{position:absolute;z-index:3;left:12px;bottom:12px;display:flex;align-items:center;gap:7px;background:rgba(13,17,23,.84);color:#fff;border:1px solid rgba(255,255,255,.18);border-radius:999px;padding:5px 12px 5px 7px;backdrop-filter:blur(3px);line-height:1}
.rank-badge .num{display:flex;align-items:center;justify-content:center;width:24px;height:24px;border-radius:50%;background:#fff;color:#0d1117;font:800 13px/1 inherit}
.rec.lead .rank-badge .num{background:var(--accent);color:#fff}
.rank-badge .lbl{font:700 10.5px/1 inherit;letter-spacing:.05em;text-transform:uppercase;color:#e6edf3}
.proof-verdict{position:absolute;z-index:3;left:12px;top:42px}
.proof-verdict .verdict{box-shadow:0 1px 6px rgba(0,0,0,.25)}
.proof-src{position:absolute;z-index:3;right:12px;bottom:12px;background:rgba(255,255,255,.92);color:var(--mut);border:1px solid var(--line);border-radius:6px;padding:4px 8px;font:700 10px/1.3 inherit;letter-spacing:.02em}
.proof-src b{color:var(--ink)}
/* GRAFT 3 — honest crop tag */
.crop-tag{position:absolute;z-index:3;right:12px;top:42px;font:700 9.5px/1 inherit;letter-spacing:.03em;color:#fff;background:rgba(31,35,40,.78);border-radius:5px;padding:4px 7px}
/* GRAFT 4 (JS-optional) — Expand button, hidden until .has-js; proof already legible without it */
.zoombtn{display:none;position:absolute;z-index:3;right:12px;top:38px;border:1px solid rgba(255,255,255,.35);background:rgba(13,17,23,.7);color:#fff;border-radius:6px;padding:4px 8px;font:700 10px/1 inherit;letter-spacing:.02em;cursor:pointer}
.has-js .zoombtn{display:inline-block}
.has-js .crop-tag{right:74px}
/* image-missing fallback */
.rec-proof.img-missing img{display:none}
.rec-proof.img-missing .frame{display:flex;align-items:center;justify-content:center;aspect-ratio:16/10;background:repeating-linear-gradient(45deg,#161b22,#161b22 12px,#1b212a 12px,#1b212a 24px);color:#9aa4af;font:600 12px/1.5 inherit;text-align:center;padding:18px}
.rec-proof .fallback{display:none}
.rec-proof.img-missing .fallback{display:block}
.rec-proof.img-missing .browserbar{display:none}

/* decision side */
.rec-body{padding:16px 18px 15px;display:flex;flex-direction:column;min-width:0}
.rec.lead .rec-body{padding-top:18px}
.rec-body h3{margin:0 0 5px;font-size:17px;line-height:1.25;letter-spacing:-.01em}
.rec.lead .rec-body h3{font-size:19px}
.rec-what{margin:0 0 12px;color:var(--mut);font-size:13.5px;line-height:1.5}
.rec-what b{color:var(--ink)}
.rec-what code{background:var(--soft);border:1px solid #cfe2fb;border-radius:4px;padding:1px 5px;font-size:12px}
.chips{display:flex;flex-wrap:wrap;gap:7px;margin-bottom:10px}
.ev-note{font-size:12px;color:var(--mut);margin:-2px 0 11px}
.ev-note b{color:var(--ink)}
/* GRAFT 2d — labeled SKIP-IF row, bottom-pinned so cards align */
.skiprow{margin-top:auto;padding-top:11px;border-top:1px dashed var(--line);display:flex;gap:8px;align-items:baseline}
.rec.lead .skiprow{border-top-color:#bcd6f5}
.skiprow .lbl{font:700 9.5px/1 inherit;letter-spacing:.04em;text-transform:uppercase;color:var(--single-fg);background:var(--single-bg);border:1px solid var(--single-bd);border-radius:5px;padding:4px 7px;whit

…(truncated)